Machine learning models are increasingly used to forecast weather patterns, predict molecular properties, and propose new materials, often by representing these systems as graphs: networks of connected parts. While these models can produce highly accurate predictions, scientists frequently cannot explain how a model arrived at its answer. A model might land on the correct result while relying on entirely the wrong reasoning, a gap that becomes critical the moment these tools are used to inform real scientific conclusions.

Lucic’s project sets out to open these “black boxes.” By tracing the exact computational steps inside graph-based machine learning models, the research aims to give scientists the tools to verify whether a model’s predictions actually follow real physical laws, rather than relying on the model’s output on faith. That kind of verification is what determines whether graph-based AI can be trusted as a genuine tool for scientific discovery, rather than a black box that happens to get the right answer.

Lucic is one of 28 UvA and Amsterdam UMC researchers to receive a Veni grant in this round. The Veni is part of NWO’s Talent Programme, an individual grant aimed at adventurous, talented researchers early in their careers, giving them the space to develop their own research ideas. NWO awarded 200 research projects nationwide in this funding round.